AUTOMATIC BUILDING EXTRACTION FROM HIGH RESOLUTION AERIAL
IMAGES USING ACTIVE CONTOUR MODEL
Salman Ahmady, Hamid Ebadi, M.J. Valadan Zouj, Hamid Abrishami Moghaddam
Photogrammetry, K.N. Toosi university of Technology, Dept of Photogrammetry and Remote Sensing, Tehran, Iran salman.ahmady@gmail.com
ABSTRACT:
Various governmental organizations need accurate, correct and up to date information for optimization of resource and service
management. In this issue, geospatial information is very important. Geospatial information as essential part of Geospatial
Information System (GIS) has important role in performance of civil projects, urban service management. Using conventional
surveying methods for producing geospatial data require a lot of cost and time. Thus, utilization of modern methods in production
and updating of this kind of data is necessary.Photogrammetry and Remote Sensing are methods that produce geospatial data in
extensive area with acceptable accuracy. In various countries of the world, many researches have been carried out and many
algorithms have been introduced in order to decrease human operation in automatic feature extraction of satellite images. Building is
one of the features that take the maximum of time and cost of feature extraction due to its abundance in urban area. As a result, on
access to a model or algorithm of automatic or semi-automatic extraction of this feature not only minimizes human role in producing
large scale maps but also has a dramatic effect on time and cost of the project. The aim of this paper is automatic extraction of
boundary of this feature from high resolution aerial images in a way that its output is a vector map that needs the least editing in GIS.
the main goal of this research is to introduce a method based on active contour model that the initialization stage of algorithm can be
carried out automatically and active contour be ultimately optimized in building extraction. A new model is also suggested for
automatic detection and extraction of boundary of buildings. New model of active contour can detect and extract boundary of
building very accurately compared to classical model of active contour model and avoid detection of the boundary of features that are
in neighbor of buildings such as streets and trees. The result of applying this model shows that the active contour model works better
than other models of detection and extraction of building boundaries in urban area.
information such as Lidar data. Halla et al. (1999) have
extracted location of buildings from image by using
classification algorithms and height data. Zimmermann et al.
(2000) produced DSM data from stereo images. Then they
detected building's roof model by applying slope and aspect
operators. Also in the (Zhao et al, 2000, Jin et al, 2005)height
data and morphological operators have been used for extraction
of buildings.
1. INTRODUCTION
Photogrammetry and Remote Sensing are methods that produce
geospatial data in extensive area with acceptable accuracy.
Furthermore, project cost and time of using these methods is
lower than other production methods of geospatial information.
Because of the fact that nowadays, spatial and spectral
resolution of satellite sensors has been improved and high
quality images such as Ikonos and Quickbird are accessed to
provide needs of civil users, utilization of these images for
providing and updating of maps and geospatial information has
been approved all over the world.
In some researches, different active contour models such as
Snake model have been employed for building boundary
recognition. In (Peng et al, 2004) a specific Snake model has
been introduced for building boundaries from aerial images with
a new energy function. The paper of) Mayunga et al,
2005)proposed a semi automatic algorithm for building
extraction from Quickbird images. In this paper within the
boundary of each building, a point is selected. Then the initial
curves of the model are produced and the accurate boundaries of
buildings are detected in the iterative procedure. In the another
paper (Guo et al, 2002) at first the approximate position of
building's boundary is extracted from Lidar data and then Snake
Model extract precise boundary of each building.
Building is one of the features made by human that takes the
maximum of time and cost of feature extraction due to its
abundance in urban area. As a result, on access to a model or
algorithm of automatic or semiautomatic extraction of this
feature not only minimizes human role in producing large scale
maps but also has a dramatic effect on time and cost of the
project. Finding such an algorithm can also be efficient in
automatic extraction of other features similar to buildings.
Because of irregular structure and closeness of buildings in
urban area, the maximum researches in domain of automatic
building extraction from high resolution aerial and satellite
images, are done from integration of Lidar data and images
(Rottensteiner et al, 2005). For example in the research of (Sohn
et al, 2007) at first all features that have a certain height from
earth are recognized. After that by using NDVI index and other
information, separate the buildings from other features. Then
the sharp edges detect and a polygon fits to the close edges as a
building boundary. In other researches such as (Weidner et al,
1995, Baillard et al, 1999, Schenk et al, 2002, Guo et al, 2002,
Rottensteiner et al, 2005) the position of buildings have been
detected by integration of aerial and satellite images into height
Other researches have been used artificial intelligence such as
nero-fuzzy system for recognition and detection buildings from
other object in aerial and satellite images (Samadzadegan et al,
2005).
In this paper a model has been introduced for building detection
from aerial images using active contour models based on level
set formulation.
2. DESCRIPTION OF ACTIVE CONTOUR MODEL
453
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
The active contour model that is used in this paper is proposed
by Chan and Vese (2001) based on Mamford-Shah image
segmentation equation. In this model image is segmented to
regions that the value of all pixels within each region have
maximum homogeneity and similarity. The most important
advantage of this active contour model is the ability of them to
extract objects without obvious edges from images. Another
advantage of this model is no sensitivity to the noise.
equations:
Active contour model based on level set formulation is
described by this energy function:
In the last step, the equation of curve evolution is:
F1 (C ) + F2 (C ) = ∫
u 0 − c1 dx + ∫
2
inside ( C )
u 0 − c 2 dx
2
outside ( C )
Where C is the curve of active contour model, and
pixel's value of the image and c1 ,
pixel's value inside and outside of C.
c2
c1 (φ ) =
u 0 H (φ )dx
Ω
∫ H (φ )dx
∫ u (1 − H (φ )dx
(φ ) =
∫ (1 − H (φ )dx
Ω
c2
Ω
(6)
0
Ω
⎡
⎤
⎛ ∇φ ⎞
∂φ
⎟ −ν − λ1 (u 0 − c1 )2 + λ2 (u 0 − c2 )2 ⎥ = 0
= δ (φ ).⎢ μ.div⎜⎜
⎟
∂t
∇
φ
⎝
⎠
⎣⎢
⎦⎥
(1)
u0
∫
Finally the above curve evolution equation is solved based on
finite difference methods and the position of curve is improved
in the iteration manner.
is
are the average of
3. AUTOMATIC BUILDING EXTRACTION USING
ACTIVE CONTOUR MODEL
If the cure C is inside of target object then F1 (C ) ≈ 0 and
F2 (C ) > 0 and in case that curve C is outside of the target object
then F1 (C ) > 0 and F2 (C ) ≈ 0 and finally if curve C is inside
and outside of target object then F1 (C ) > 0 and F2 (C ) > 0 .
After minimization of function of equation (1), the curve C is
fitted to the boundary of target object and this relation is
obtained:
Our active contour model for building extraction is described in
section 2. Our building extraction method is made up four major
steps. In the first stage the input image is smoothed using a
Gaussian kernel. In next step a point inside a building is
introduced to model as a training data. Then the proposed active
contour model is implemented on the smoothed image and
boundaries of building are detected. Finally the accuracy of
detected buildings is evaluated. Figure 1 demonstrates the
diagram of steps of proposed model.
inf c {F1 (C ) + F2 (C )} ≈ 0 ≈ {F1 (C0 ) + F2 (C0 )}
For regularization of equation (1), two terms are added to it and
the equation (2) is obtained:
F (C , c1 , c2 ) = μ .(length(C ) ) +ν .area(inside(C ) ) +
p
λ1 ∫
u0 − c1 dx + λ2 ∫
2
inside ( C )
(2)
u0 − c2 dx
2
outside ( C )
Where μ ≥ 0 , ν ≥ 0 and λ1 , λ2 > 0 are constant
parameters. The equation (2) has a level set formulation that
was introduced in (Chan et al, 2001). For this purpose they
defined two additional functions:
d
z≥0
⎧1,
δ ( z) = H ( z)
H ( z) = ⎨
dz
(3)
0
,
z
<
0
⎩
Based on equation (3) the definition of each part of equation (2)
is:
Figure 1: diagram of building extraction steps
Our model is tested on an aerial image from Gachsaran city in
Iran. The buildings in this image are dense and attached
together. Figure 2 shows the image of test region.
length{φ = 0} = ∫ ∇H (φ ) dx = ∫ δ (φ ) ∇φ dx
Ω
area{φ ≥ 0} = ∫ H (φ )dx
Ω
∫φ
∫φ
(4)
≥0
u 0 − c1 dx = ∫ u 0 − c1 H (φ )dx
<0
u 0 − c 2 dx = ∫ u 0 − c 2 (1 − H (φ ) )dx
2
2
Ω
2
2
Ω
In the above equations Ω shows the entire of image domain.
With regarding the equation (4), the level set definition of
equation (1) is (Chan et al, 2001):
F (φ , c1 , c 2 ) = μ
(∫
Ω
)
p
∇H (φ ) dx + ν ∫ H (φ )dx +
Ω
λ1 ∫ u 0 − c1 H (φ ) dx + λ2 ∫ u 0 − c 2 (1 − H (φ ) )dx
2
Ω
2
Ω
(5)
In this equation, c1 and c 2 are the mean of pixel's value if
φ ≥ 0 and φ < 0 respectively and they are obtained from these
Figure 2: Gachsaran city as the test region
454
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
At first step input image is smoothes with a Gaussian kernel:
noise. Previous models in building extraction don't have this
accuracy in dens, irregular an attached buildings urban regions.
A problem of our model is in detection of buildings that have
similar spectral information with other features such as streets.
I g = Gσ * I
Where I g , I and Gσ are the smoothed image, input image and
Gaussian kernel respectively. By introducing a point inside a
building as sample data, the model runs and the initial curves
are generated automatically which they are a series of regular
circles (figure 3).
4. ACCURACY ASSESSMENT OF THE MODEL
In this paper, we use the McKeon's shape accuracy factor for
evaluation of the model. In this relation the area of buildings in
ground truth is compared with the area of buildings that is
detected by model.
shape accuracy = (1 −
A− B
A
) *100
Where A and B are area of a building in ground truth and area
of its corresponding detected building respectively. Table 1
shows the results of the model.
shape accuracy(%)
Figure 3: initial curves generation
max
min
mean
80
60
75
Table 1: the shape accuracy obtained from the model
5. CONCLUSION
In this paper, an optimum model of Active contour models is
utilized for automatic building extraction from aerial images.
This model does not need introduction of initial curves near
edge of buildings. Results of our model applied to aerial images
show suitable results especially in dense and irregular urban
areas.
Step 1
REFERENCES
Rottensteiner, F, Trinder, J, Clode, S, Kubik, K, Using the
Dempster–Shafer method for the fusion of LIDAR data and
multispectral images for building detection, Information Fusion
6 (4), 283–300, 2005
Sohn, G, Dowman, I, Data fusion of high-resolution satellite
imagery and LiDAR data for automatic building extraction,
ISPRS Journal of Photogrammetry & Remote Sensing 62, 43–
63, 2007
Step 2
Guo, T, Yasuoka,. Y,. Snake-based approach for building
extraction from high-resolution satellite images and height data
in urban areas, Proceedings of the 23rd Asian Conference on
Remote Sensing, November 25–29, Kathmandu. 7, 2002
Schenk, T, Csatho, B, Fusion of LiDAR data and aerial imagery
for a more complete surface description. International Archives
of Photogrammetry & Remote Sensing and Spatial Information
Sciences 34 (Part 3), 310–317, 2002
Weidner, U, Forstner, W, Towards automatic building
reconstruction from high resolution digital elevation models,
ISPRS Journal of Photogrammetry and Remote Sensing 50 (4),
38–49, 1995
Baillard, C, Maitre, H, 3D reconstruction of urban scenes from
aerial stereo imagery: A focusing strategy, Computer Vision
Step 3
Figure4. Curve position in three iterations todetect buildings
boundary
The result of implemented model shows that the boundaries of
buildings are detected accurately and model is not sensitive to
455
The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences. Vol. XXXVII. Part B3b. Beijing 2008
and Image Understanding 76 (3), 244–258, 1999
Halla, N C.Brenner, Extraction of building and trees in urban
environment, ISPRS Journal of Photogrammetry & Remote
Sensing 54, 130-137, 1999
Zimmermann, P, A new framework for automatic building
detection analyzing multiple cue data, International Archives of
Photogrammetry and Remote Sensing, Vol.XXXIII, part B3,
2000
Zhao, B, Trinder, J.C, Intgrated-Approach-Based automatic
building extraction, International Archives of Photogrammetry
and Remote Sensing, Vol.XXXIII, part B3, 2000
Jin, X, Davis, C.H, Automatic building extraction from high
resolution satellite imagery in urban areas using structural,
contextual, and spectral information, EURASIP Journal of
Applied Signal Processing 14, 2196-2206, 2005
Peng, J, Zhang, D, Liu, Y, An improved snake model for
building detection from urban aerial images, Pattern recognition
Letters, 2004, 587-595
Mayunga , S.D, Zhang, Y, Coleman, D.J, Semi-Automatic
building extraction utilizing Quickbird imagery, IAPRS, Vol.
XXXVI, Part 3/W24, 29-30, 2005
Samadzadegan, F, Azizi, A, Hahn, M, Lucas, C, Automatic 3D
object recognition and reconstruction based on neuro-fuzzy
modeling, ISPRS Journal of Photogrammetry & Remote
Sensing 59, 255– 277, 2005
Chan, T.F, Vese, L.A, Active contours without edges, IEEE
Trans. Image Processing 10, 266–277, 2001
456